1 Times of Newton, Euler, Lagrange and Poinaré
How are the trajectories of three point masses and that are attracted each other by Newton’s gravitational law? This socalled “threebody problem” can be traced back to Newton[31] in 1687. According to Newtion’s second law and gravitational law, the related governing equations about body problem read
(1) 
where and
are the position vector and mass of the
thbody, denotes the time, respectively, with a given initial position and velocity , i.e.(2) 
Here the dot denotes the derivative with respect to . Note that , and are dimensionless using a characteristic length , a characteristic mass and the Newton’s gravitational constant . If the trajectory of each body at exactly returns its initial status, say,
(3) 
one has a periodic solution of the body problem. The famous threebody problem corresponds to = 3.
For the twobody problem, corresponding to = 2, Newton gave a closedform periodic solution. However, for a threebody problem ( = 3 ), it becomes extremely difficult to find periodic orbits: no periodic orbits had been found until Euler reported one in 1740 and Lagrange published one in 1772, both for a triple system with three equal masses. However, according to Montgomery’s topological method[29]
of classifying periodic orbits of threebody systems, they belong to the
same family, namely the “Euler  Lagrange family”. Thereafter, no new periodic orbits were reported in about two centuries.Why is it so difficult to find periodic orbits of threebody systems? The mysterious reasons were revealed by Poinaré[33] in 1890, who proved that, unlike the twobody problem that is integrable and thus its solutions is completely understood, the threebody problem is not integrable: its first integral does not exist at all. This well explains why only the Euler  Lagrange family (in closed form) were found in more than two hundred years, since closedform orbits do not exist at all in general. It implies that one generally had to use numerical algorithms to solve threebody problem, but unfortunately the electronic computer was even not invented in the times of Poinaré[33]. In addition, Poinaré[33] also found that trajectories of threebody system are generally rather sensitive to initial conditions (i.e. the butterflyeffect), which leaded to the foundation of a new field of modern science, i.e. chaotic dynamics. Nowadays, it is wellknown that, due to the famous butterflyeffect, i.e. a hurricane in North America might be created by a flapping of wings of a distant butterfly in South America several weeks earlier, it is very difficult to gain reliable trajectories of chaotic systems, especially in a long interval of time. All of these explain why periodic orbits of threebody problem are so difficult to obtain and why it becomes one of the oldest open question in science.
In most cases, trajectories of threebody system are chaotic, i.e. nonperiodic, as discovered by Poinaré[33] in 1890 and confirmed again by Lorenz[26] in 1963 and Stone & Leigh[36] in 2019. However, in some special cases, there indeed exist periodic orbits, i.e. the three bodies exactly return to their initial positions and initial velocities after a period . The periodic orbits of the threebody problems are very important, since they are “the only opening through which we can try to penetrate in a place which, up to now, was supposed to be inaccessible”, as pointed out by Poinaré[33]. However, the question is: how can we find them effectively?
2 Times of supercomputer
The excellent work of Poinaré[33] made a historical turning point of searching for periodic orbits of threebody system. The nonexistence of the uniform first integral of triple system reveals the impossibility of finding closedform analytic solutions of threebody problems in general cases: it clearly indicates that we mostly had to (i.e. must) use numerical algorithms to solve this problem. This was indeed a revolutionary contribution of Poinaré[33] with great foresightness at that times when there was even no electronic computers at all: about a half century later, Turing[39, 40]
published his epochmaking papers that became the foundation of modern computer and artificial intelligence. Thanks to Von Neumann
[41], who proposed the socalled Von Neumann  Machine for modern computer, also to Jack S. Kilby, who won the Nobel prize in Physics 2000 for his taking part in the invention of the integrated circuit, and to many scientists, mathematicians, engineers and so on whose names we have no space to mention here, the performance of supercomputer becomes better and better, together with more and more advanced numerical algorithms. This provides us a strong support of hardware and software for discovering new periodic orbits of threebody problem now.With the ceaseless progress of electronic computers, more and more researchers followed the way, which Poinaré[33] suggested, to numerically search for periodic orbits of threebody problem. According to Montgomery’s topological method[29] of classifying periodic orbits of threebody systems, only three families of periodic orbits were found in 300 years after Newton, i.e. (1) the socalled “Euler  Lagrange family” found by Euler in 1740 and Lagrange in 1772; (2) the socalled “BHH family” for three equal masses, numerically found by Broucke[6, 5] in 1975, Hadjidemetriou[9] in 1975 and Hénon[11] in 1976, respectively; (3) the socalled “figureeight family” for three equal masses numerically found by Moore[30] in 1993, until 2013 when Šuvakov and Dmitrašinović[37] numerically gained 11 new families of periodic orbits of triple system with three equal masses. All of these suggest that finding periodic orbits of threebody problem by means of numerical methods should be a correct way.
Note that, among the three families of periodic orbits mentioned above, the BHH family and figure8 family were found by numerical methods respectively in 1975s and 1993, when the performance of computer might be not good enough. However, in 2010s we had supercomputer with peak performance about 1,000 petaflops, i.e. several billion billion fundamental calculations per second, together with lots of powerful numerical algorithms. What prevents us from effectively finding thousands of new families of periodic orbits of threebody problem though we have so powerful supercomputer ?
The key of finding periodic orbits is to gain reliable computergenerated trajectories of threebody system under arbitrary initial conditions in a long enough interval of time. However, as discovered by Poinaré[33] and rediscovered by Lorenz[26], computergenerated trajectories of chaotic systems are sensitive to initial conditions, i.e. the famous butterflyeffect. In other words, a tiny difference on initial conditions might lead to a huge deviation of computergenerated simulation after a long time. In addition, Lorenz[27, 28] further found that computergenerated trajectories of chaotic systems are also sensitive to algorithms: different numerical algorithms might give distinctly different computergenerated trajectories of chaotic systems after a long time. This is indeed a great obstacle! This kind of sensitive dependence on numerical algorithm (SDNA) for a chaotic system had been also observed and reported by many researchers[17, 18, 38], which however unavoidably leaded to some intense arguments[42]: some researchers even suggested that “all chaotic responses should be simply numerical noises” and might “have nothing to do with differential equations”. Thus, by means of traditional numerical algorithms (mostly in double precision), it is rather difficult to gain reliable/convergent computergenerated trajectories of threebody system under arbitrary initial conditions in a long enough interval of time.
To overcome this obstacle, Liao[24] proposed the socalled “clean numerical simulation” (CNS) for chaotic systems. Unlike other traditional numerical algorithms, which mostly use double precision, the CNS can greatly reduce not only truncation error but also the roundoff error to keep the total “numerical noises” in such a required tiny level that a reliable (or convergent) computergenerated simulation can be obtained in a long enough interval of time. In the frame of the CNS, the truncation error is decreased by means of numerical algorithms at high enough order in time and space, and the roundoff error is reduced by using multipleprecision with many enough digits for all parameters and variables. The CNS has been successfully applied to many chaotic systems, such as Lorenz equation, chaotc threebody systems, some spatiotemporal chaotic systems and so on[25, 23, 22, 12]. As reported by Hu and Liao[12], the use of double precision might lead to huge deviations of computergenerated simulations of spatiotemporal chaos even in statistics, not only quantitatively but also qualitatively, particularly in a long interval of time. This indicates that we must be very careful to numerically simulate chaotic systems. Fortunately, the CNS can provide us a guaranteed tool to gain reliable/convergent trajectories of chaotic systems (such as threebody systems with arbitrary initial conditions) during a long enough time.
Using the CNS as an integrator of the governing equations and combining the grid search method and the NewtonRaphson method[7, 16, 1], Li and Liao[21] in 2017 successfully found 695 families of periodic planar collisionless orbits of threebody systems with three equal masses and zero angular momentum, including the figureeight family, the 11 families found by Šuvakov and Dmitrašinović in 2013, and besides more than 600 new families that have never been reported. Similarly, Li, Jing and Liao[19] further found 1349 new families of periodic planar collisionless orbits of the threebody system with only two equal masses. In 2020, starting from a known periodic orbit with three equal masses and using the CNS to integrate the governing equations, Li, Li and Liao[20] successfully obtained 135445 new periodic orbits with arbitrarily unequal masses by means of combining the numerical continuation method[3] and the NewtonRaphson method[7, 16, 1], including 13315 stable ones. Therefore, in only four years, using highperformance computer and our new strategy based on the CNS, we successfully increased the family number of the known periodic orbits of threebody systems by nearly four orders of magnitude! This strongly indicates that our numerical strategy is correct, powerful and rather effective for finding new periodic orbits of threebody systems. It should be emphasized that this great progress is mainly due to the use of the new numerical strategy based on the CNS, since the performance of supercomputer is good enough for threebody systems even at the beginning of the 22nd century.
3 The times of machine learning
However, it is timeconsuming to use the numerical continuation method[3] to find the 135445 periodic orbits (with unequal masses) reported by Li, Li and Liao[20]. Besides, these periodic orbits are in essence discrete, say, only for some specific values of and in an irregular domain (in case of since we use the mass of the 3rd body as a characteristic mass , without loss of generality). Can we gain a periodic orbit more efficiently for arbitrary values of masses and ? Thanks to the times of machine learning, the answer is positive and rather attractive, as described below.
For example, let us use a known BHH orbit as a starting point to illustrate this. Fiftyseven satellite periodic orbits of the BHH family of threebody systems with three equal masses were found[14] in 2016, and this number was extended to ninetynine[13] in 2020. The initial configuration of the BHH satellite periodic orbits with zero angular momentum is described by
and
where , and is the position vector, velocity vector and mass of the th body, respectively. Thus, for given and (we assume thereafter), we should determine four unknown physical variables and the period . Note that these orbits are periodic in a rotating frame of reference, say, the frame of coordinates rotates an angle in the corresponding period .
Without loss of generality, let us consider such a known BHH periodic orbit with the initial condition
the period and the rotation angel of the coordinate frame, where . Using this periodic orbit as a starting point, we first follow Li, Li and Liao[20] to obtain only 36 periodic orbits for different masses in a small domain (marked by , with the mass increment ) by means of the numerical continuation method[3] and the NewtonRaphson method[7, 16, 1]. For details, please refer to Li, Li and Liao[20]. The return distance (deviation) of these periodic orbits is defined by
in the rotating frame of reference, where is the period. Note that all of the 36 periodic orbits satisfy the criteria . Besides, they all have the same family name, which is defined by the socalled “free group element” according to Montgomery’s topological method[29].
In the next step, we use an artificial neural network[10, 8, 4, 2] (ANN) model to gain a relation between the input vector (, ) and the output vector (, , ,
) of the periodic orbits. The ANN model we used here consists of multiple fully connected layers, say, one input layer, six hidden layers and one output layer. The number of neurons for the input layer, the hidden layers and the output layer is 2, 1024 and 4, respectively, as shown in Figure
1. We use the optimization algorithm AMSGrad[34] as an optimizer to minimize the mean square error for training the ANN model. At beginning, we use the results of the 36 known periodic orbits gained by means of the numerical continuation method[3] and the NewtonRaphson method[7, 16, 1] as the training set to train the ANN model. The trained ANN model provides us a kind of relationship (expressed by ) between and , which can be further used to predict the initial conditions , , and the period of candidates (i.e. possible periodic orbits) for various masses () outside of the original small domain . Using these predictions as the initial guesses for the NewtonRaphson method[7, 16, 1], we indeed successfully obtain the 17421 new periodic orbits with for various outside of , which are marked in green in Figure 2 and expressed by . Then, in the same way, we further use the results of all (i.e. 36 + 17421 = 17457) known periodic orbits as a training set to further train our ANN model so as to gain a better relationship between (, ) and (, , , ) for extrapolation/expansion outside of the mass domain , which gives us 11473 new periodic orbits in a even larger domain, as marked in purple in Figure 2 and expressed by . Similarly, we use the results of all (i.e. 36 + 17421 + 11473 = 28930 ) known periodic orbits as the training set to further train our ANN model so as to give a relationship between and , which however gives us only the 220 new periodic orbits outside of the mass domain , as marked in blue in Figure 2 and expressed by . This implies that we might find the nearly largest mass domain for the existence of periodic orbits with a few of same properties. As shown in Figure 2, starting from the 36 periodic orbits in the original small domain (marked in red), we totally gain the 29150 periodic orbits in the mass domain by the three times extrapolations/expansions. Finally, training our ANN model by using all (i.e. 29150) these known results as a training set, we obtain a relationship between and (, , , ), which can give predictions of periodic orbits for arbitrary values of in the accuracy level of for the return distance (deviation) . Note that the mass domain with the periodic orbits becomes larger and larger, i.e. from to to to , and the relationship between and (, , , ) becomes better and better, from to to to , implying that our ANN model becomes wiser and wiser!Obviously, the smaller the return distance (deviation) is, the more accurate the periodic orbit given by the numerical strategy. Note that exactly corresponds to a closedform periodic orbit. But unfortunately, except the Euler  Lagrange family, nearly all known periodic orbits of threebody systems were gained by numerical methods, and this fact is consistent with Poinaré’s famous proof of the nonexistence of the uniform first integral of threebody systems[33]. It is found that, since the CNS and the MP (multiple precision) are used in our numerical strategy, the return distance (deviation) of these periodic orbits can be reduced to any given value, say, the initial conditions and the period of these relatively periodic orbits can be at arbitrary level of precision, i.e. as accurate as one would like, as long as the performance of supercomputer is good enough.
The relationship between and given by the ANN model is fundamentally different from the set of the original 29150 periodic orbits in the following aspects.

For arbitrary values of , can always give a good enough prediction of the initial condition and the period of the corresponding periodic orbit, which can be used as a starting point to gain a more accurate periodic orbit with a tiny return distance (deviation) . Therefore, in theory, our ANN model can provide us with an infinite number of periodic orbits.

For each prediction of a periodic orbit given by our ANN model for arbitrary values of , we can always modify it by means of the NewtonRaphson method[7, 16, 1], until a very accurate periodic orbit with is gained. For examples, for the randomly chosen masses and , the initial condition and period of the periodic orbit predicted by is , , and . The return distance of this predicted periodic orbit is about , which can be reduced to by means of the NewtonRaphson method[7, 16, 1] and the CNS, with the initial condition and period in accuracy of hundred significant digits:
Note that the prediction given by our ANN model are in the accuracy of four significant digits. Similarly, we randomly check one thousand cases of , and found that, in every case, we indeed can always gain a periodic orbit with a quite tiny return distance (deviation) . All of these suggest that, statistically speaking, for arbitrary mass values , every prediction given by our ANN model could lead to a periodic orbit that could be in arbitrary accuracy as long as the performance of computer is good enough.

It is found that the errors of the predictions given by our trained ANN model are mostly in the level of for the initial conditions and for the period . Although we indeed can further gain a periodic orbit in accuracy of one hundred significant digits by means of the NewtonRaphson method[7, 16, 1] and the CNS, the periodic orbits predicted by our trained ANN model are good enough from practical viewpoint, since it is unnecessary to have so accurate trajectory in practice such as astronomical observation and so on. In practice, it is rather convenient to gain a periodic orbit for arbitrary masses by means of the initial conditions and period predicted by the trained ANN model, for example, as shown in Figure 3 for randomly chosen six different masses of and . Thus, our ANN model provides us great convenience in practice.

The ANN model can ceaselessly learn and thus be further modified when some new periodic orbits are gained, say, the trained ANN model could become wiser and wiser.
In summary, we illustrate that the trained ANN model can provide us an infinite number of periodic orbits for arbitrary values of in an irregular domain .
How accurate is a periodic orbit with return distance (deviation) ? Even if we use the diameter of universe = 930 light year = meter as the characteristic length , we have the corresponding inaccuracy of the dimensional initial position meter that is even smaller than the Planck length meter. Note that Planck length is a lower bound to physical proper length in any spacetime: it is impossible to measure length scales smaller than Planck length, according to Padmanabhan[32]. Besides, it should be emphasized that all of these periodic orbits are stable, say, a tiny disturbance does not increase exponentially. So, from physical viewpoint, a stable periodic orbit with return distance (deviation) gained by numerical method is equivalent to that corresponds to an exact solution of periodic orbit in a closedform.
Therefore, the highperformance computer and the machine learning play very important role in finding periodic orbits of threebody systems with arbitrary masses. It should be emphasized here that it is Turing[39, 40] who laid the foundations of modern computer and artificial intelligence (including machine learning).
The above approach has general meaning. To show this point, let us further consider another BHH satellite periodic orbit with three equal masses , = , , , the period and the rotation angel of the reference frame (for relatively periodic orbits). At first, using this known periodic orbit as a starting point, we can obtain 36 periodic orbits with various masses in a small mass domain and (with mass increment ) by combining the numerical continuation method[3] and the NewtonRaphson method[7, 16, 1].
Similarly, we train a ANN model (with the same structure as mentioned above) by means of the initial conditions and periods of these 36 known periodic orbits and then use the trained ANN model to predict the initial conditions and periods of some candidates of possible periodic orbits outside the previous mass domain, while the initial conditions and periods of each candidate are modified by the NewtonRaphson method[7, 16, 1] so as to confirm whether or not the candidate is a periodic one, say, its return distance can be reduced to the tiny level of . The same process repeats 11 times, while the number of the known relatively periodic orbits becomes more and more, the mass domain becomes larger and larger, and the ANN model becomes wiser and wiser! Finally, we totally obtain 36251 relatively periodic orbits in a mass domain whose area is about 1000 times larger than the initial one and , as shown in Fig. 4.
Similarly, we further train our ANN model by using the known 36251 relatively periodic orbits. It is found that, for randomly chosen one thousand cases of arbitrary mass , our trained ANN model can always predict the initial conditions and periods that often provide us corresponding periodic orbits with a return distance in a tiny level of , and more importantly, that can be always further improved until a very small return distance is satisfied. Thus, statistically speaking, our trained ANN model can always give periodic orbits for arbitrary mass . It is found that our trained ANN model is accurate enough from practical viewpoint. For example, randomly choosing six cases with various masses and , we can use our ANN model to quickly predict their initial conditions and periods, as listed in Table 2, and give the corresponding trajectories in a satisfied level of accuracy, as shown in Fig 5. Thus, our approach based on the ANN model has indeed general meaning.
Stability is an important property of periodic orbits, because only stable triple systems can be observed. We employ a theorem given by Kepela and Simó[15] to determine the linear stability of periodic orbits of threebody problem through the monondromy matrix [35]. It is found that all relatively periodic orbits for the first case with the rotation angle are linearly stable. For the second case with the rotation angle , there are 16739 linearly stable periodic orbits among the 35895 computergenerated periodic orbits, as shown in Fig 6. It should be emphasized that we can give an ANN classifier model to classify the periodicity and stability of the orbits for arbitrary masses and . The orbits are classified into three categories: stable periodic orbits, unstable periodic orbits and nonperiodic orbits, expressed by , and
, respectively. The ANN classifier model consists of eight fully connected layers with one input layer, six hidden layers and one output layer. The numbers of neurons for input layer, hidden layers and output layer are 2, 256 and 3, respectively. Different from the abovementioned ANN regression model, the activation function of the ANN classifier model is softmax function. The loss function is cross entropy. For each case, the whole dataset is randomly divided into two sets, the training set (90%) and test set (10%). The
of the entire data are used to train the ANN classifier model, and the remaining is used to verify the generalization capability of the model. We train the model twenty thousand epochs and choose the model with maximum accuracy on the training set. The accuracy of the ANN classifier model is about . Finally, we use the ANN classifier model to predict the periodicity and stability of orbits for any given masses and in each case.4 A roadmap of searching for periodic orbits of threebody problem
The successful examples mentioned above suggest us a general road map for finding new periodic orbits of threebody system (in case of ) with the same “free group element” (word) given by Montgomery’s topological method[29]:

For a threebody system with three or two equal masses, first find candidates of the initial conditions for possible periodic orbits by means of the grid search method, and then modify these candidates by means of the NewtonRaphson method[7, 16, 1], until a satisfied periodic orbit with a tiny enough return distance (deviation) is obtained;

Given a known periodic orbit, use it as a starting point to gain a few of new periodic orbits with various masses in a small domain of mass by combining the numerical continuation method[3] and the NewtonRaphson method[7, 16, 1]. The initial conditions and periods of all these known periodic orbits form a training set for a ANN model.

For a given training set in a mass domain of , train the ANN model to predict initial conditions and periods so that some new periodic orbits outside of the previous mass domain of could be found by modifying these predictions via the NewtonRaphson method[7, 16, 1]. Then, combining the results of these new periodic orbits with the previous training set, we further have a new training set with more elements, which could further provide us some new periodic orbits in a even larger mass domain of in a similar way, outside of the previous one. The same process can repeat again and again, so that more and more periodic orbits are found in a larger and larger mass domain of , and the ANN model becomes wiser and wiser, until quit few or no new periodic orbits can be found in a larger domain of . Finally, we have a trained ANN model of all periodic orbits in the final mass domain .

Randomly choose hundreds or thousands of arbitrary masses . For each case, check whether or not the trained ANN model could give an accurate enough prediction, and in addition whether or not the corresponding return distance (deviation) could be indeed reduced to . If yes, then statistically speaking, the ANN model can mostly provide a good enough prediction of periodic orbits in the domain , from practical viewpoint.
Note that it is better to use the CNS as integrator so as to guarantee the reliability and convergence of computergenerated trajectories of chaotic threebody system under arbitrary initial conditions in a required long interval of time. Besides, the periodicity and stability of orbits for arbitrary masses and can be well predicted by an ANN classifier model.
5 Concluding remarks and discussions
The famous threebody problem can be traced back to Newton in 1687, but quite few families of periodic orbits were found in 300 years thereafter. As proved by Poincarè, the first integral does not exist for threebody systems, which implies that numerical approach had to be used in general. In this paper, we propose an effective approach and roadmap to numerically gain planar periodic orbits of threebody systems with arbitrary masses by means of machine learning based on an artificial neural network (ANN) model. Given any a known periodic orbit as a starting point, this approach can provide more and more periodic orbits (of the same family name) with variable masses, while the mass domain having periodic orbits becomes larger and larger, and the ANN model becomes wiser and wiser. Finally we have an ANN model trained by means of all obtained periodic orbits of the same family, which provides a convenient way to give accurate enough periodic orbits with arbitrary masses for physicists and astronomers. In addition, the periodicity and stability of orbits for arbitrary masses can be well predicted by an ANN classifier model.
It must be emphasized that high performance computer and artificial intelligence (including machine learning) play important roles in solving periodic orbits of triple systems. Today, nothing can prevent us from obtaining massive periodic solutions of threebody problem. This is due to the great contributions of some great mathematicians, scientists and engineers in more than three hundred years! Especially, it is Poinaré[33] who made a historical turning point by proving the nonexistence of the uniform first integral of triple system, which implies that we had to use numerical approach in general. It is Turing[39, 40] who published his epochmaking papers that became the foundation of modern computer and artificial intelligence. It is Von Neumann[41] who proposed the socalled “Von Neumann  Machine” for modern computer. It is Jack S. Kilby, the winner of Nobel prize for physics in 2000, who took part in the invention of the integrated circuit. And so on. The famous threebody problem might be an excellent example to illustrate the importance of inventing new tools for human being to better understand and explore the nature.
The approach and roadmap mentioned in this paper has general meanings. Note that thousands of families of periodic orbits of triple systems with three or two equal masses have currently been found (for example, please visit the website https://github.com/sjtuliao/threebody). Using any of them as a starting point, we can similarly gain a trained ANN model to give accurate enough predictions of periodic orbits of the same family of triple system in a corresponding mass domain. All of these might form a massive data base for periodic orbits of triple systems, which should be helpful to enrich and deepen our understandings about the famous threebody systems. Besides, hopefully, some physical laws, such as generalied Kepler’s law[21] for periodic orbits of triple systems with arbitrary masses, could be found in future by analysing these massive data by means of machine learning.
Note also that nearly all known periodic orbits of triple systems are planar, i.e. twodimensional. In theory, the same ideas mentioned in this paper can be used to search for periodic orbits of triple systems in three dimensions.
Reference
References
 [1] (201107) Computing periodic orbits with arbitrary precision. Phys. Rev. E 84, pp. 016701. Cited by: §2, item (B), item (C), §3, §3, §3, §3, item (1), item (2), item (3).
 [2] (2018) Stateoftheart in artificial neural network applications: a survey. Heliyon 4 (11), pp. e00938. Cited by: §3.
 [3] (2003) Introduction to numerical continuation methods. Vol. 45, SIAM. Cited by: §2, §3, §3, §3, §3, item (2).
 [4] (1995) Survey and critique of techniques for extracting rules from trained artificial neural networks. Knowledgebased systems 8 (6), pp. 373–389. Cited by: §3.
 [5] (1975) Periodic orbits in the planar general threebody problem. Celestial mechanics 11 (1), pp. 13–38. External Links: ISSN 15729478 Cited by: §2.
 [6] (1975) On relative periodic solutions of the planar general threebody problem. Celestial Mechanics 12 (4), pp. 439 – 462. External Links: ISSN 15729478 Cited by: §2.
 [7] (1995) Methods for locating periodic orbits in highly unstable systems. Journal of Molecular Structure: THEOCHEM 341 (1), pp. 91 – 100. Note: External Links: ISSN 01661280 Cited by: §2, item (B), item (C), §3, §3, §3, §3, item (1), item (2), item (3).
 [8] (2003) Review and comparison of methods to study the contribution of variables in artificial neural network models. Ecological modelling 160 (3), pp. 249–264. Cited by: §3.
 [9] (1975) The stability of periodic orbits in the threebody problem. Celestial Mechanics 12 (3), pp. 255–276. External Links: ISSN 15729478 Cited by: §2.
 [10] (1995) Fundamentals of artificial neural networks. MIT press. Cited by: §3.
 [11] (1976) A family of periodic solutions of the planar threebody problem, and their stability. Celestial mechanics 13 (3), pp. 267–285. External Links: ISSN 15729478 Cited by: §2.
 [12] (2020) On the risks of using double precision in numerical simulations of spatiotemporal chaos. Journal of Computational Physics 418, pp. 109629. Cited by: §2.
 [13] (2020) A guide to hunting periodic threebody orbits with nonvanishing angular momentum. Computer Physics Communications 250, pp. 107052. Cited by: §3.
 [14] (201602) Angular momentum and topological dependence of Kepler’s third law in the BrouckeHadjidemetriouHénon family of periodic threebody orbits. Phys. Rev. Lett. 116, pp. 064301. Cited by: §3.
 [15] (2007) Computer assisted proofs for nonsymmetric planar choreographies and for stability of the Eight. Nonlinearity 20 (5), pp. 1241. Cited by: §3.
 [16] (2002) On the numerical continuation of periodic orbits  an intrinsic, 3dimensional, differential, predictorcorrector algorithm. Astronomy and Astrophysics 389 (2), pp. 692–701. Cited by: §2, item (B), item (C), §3, §3, §3, §3, item (1), item (2), item (3).

[17]
(2000)
Computational uncertainty principle in nonlinear ordinary differential equations (i)
. Science in China 43, pp. 449–460. Cited by: §2.  [18] (2001) Computational uncertainty principle in nonlinear ordinary differential equations. Science in China Series E: Technological Sciences 44 (1), pp. 55–74. Cited by: §2.
 [19] (2018) Over a thousand new periodic orbits of a planar threebody system with unequal masses. Publications of the Astronomical Society of Japan 70 (4), pp. 64. Cited by: §2.
 [20] (2021) One family of 13315 stable periodic orbits of nonhierarchical unequalmass triple systems. Science China Physics, Mechanics & Astronomy 64 (1), pp. 219511. Cited by: §2, §3, §3.
 [21] (2017) More than six hundred new families of Newtonian periodic planar collisionless threebody orbits. Science China Physics, Mechanics & Astronomy 60 (12), pp. 129511. Cited by: §2, §5.
 [22] (2015) On the inherent selfexcited macroscopic randomness of chaotic threebody systems. Int. J. Bifurcation & Chaos 25, pp. 1530023 (11 page). Cited by: §2.
 [23] (2014) On the mathematically reliable longterm simulation of chaotic solutions of Lorenz equation in the interval [0,10000]. Sci. China  Phys. Mech. Astron. 57, pp. 330 – 335,. Cited by: §2.
 [24] (2009) On the reliability of computed chaotic solutions of nonlinear differential equations. Tellus A 61 (4), pp. 550–564. Cited by: §2.
 [25] (2014) Physical limit of prediction for chaotic motion of threebody problem. Communications in Nonlinear Science and Numerical Simulation 19 (3), pp. 601–616. Cited by: §2.
 [26] (1963) Deterministic nonperiodic flow. Journal of the Atmospheric Sciences 20 (2), pp. 130–141. Cited by: §1, §2.
 [27] (1989) Computational chaosa prelude to computational instability. Physica D: Nonlinear Phenomena 35 (3), pp. 299–317. Cited by: §2.
 [28] (2006) Computational periodicity as observed in a simple system. Tellus A 58, pp. 549 – 557. Cited by: §2.
 [29] (1998) The Nbody problem, the braid group, and actionminimizing periodic solutions. Nonlinearity 11 (2), pp. 363. Cited by: §1, §2, §3, §4.
 [30] (199306) Braids in classical dynamics. Phys. Rev. Lett. 70, pp. 3675–3679. Cited by: §2.
 [31] (1687) Philosophiæ naturalis principia mathematica (Mathematical Principles of Natural Philosophy). London: Royal Society Press. Cited by: §1.
 [32] (1985Nov.) Physical significance of Planck length. Annals of Physics 165, pp. 38 – 58. Cited by: §3.
 [33] (1890) Sur le probléme des trois corps et les équations de la dynamique. Divergence des séries de m. Lindstedt. Acta Math. 13, pp. 1–270. Cited by: §1, §1, §2, §2, §2, §3, §5.
 [34] (2019) On the convergence of Adam and beyond. arXiv preprint arXiv:1904.09237. Cited by: §3.
 [35] (2002) Dynamical properties of the figure eight solution of the threebody problem. Celestial Mechanics: Dedicated to Donald Saari for His 60th Birthday: Proceedings of an International Conference on Celestial Mechanics, December 1519, 1999, Northwestern University, Evanston, Illinois 292, pp. 209. Cited by: §3.
 [36] (2019) A statistical solution to the chaotic, nonhierarchical threebody problem. Nature 576 (7787), pp. 406–410. Cited by: §1.
 [37] (201303) Three classes of Newtonian threebody planar periodic orbits. Phys. Rev. Lett. 110, pp. 114301. Cited by: §2.
 [38] (2007) Time step sensitivity of nonlinear atmospheric models: numerical convergence, truncation error growth, and ensemble design. Journal of the Atmospheric Sciences 64 (1), pp. 175–189. Cited by: §2.
 [39] (1936) On computable numbers, with an application to the Entscheidungs problem. Proc. London Maths. Soc. 42 (), pp. 230–265. Cited by: §2, §3, §5.
 [40] (1950) Computing machinery and intelligence. Mind 50 (), pp. 433–460. Cited by: §2, §3, §5.
 [41] (1958) The computer and the brain. Yale University Press. Cited by: §2, §5.
 [42] (2008) Comment on“computational periodicity as observed in a simple system” by Edward N. Lorenz (2006). Tellus A 60, pp. 803 – 805. Cited by: §2.
Comments
There are no comments yet.